GAN-Based SEM Image Simulation for Broader Denoising Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing denoising models for SEM images require extensive training data from patterned substrates, leading to impractical metrology time and computing resources, with limited pattern coverage and frequent retraining needs.
Innovation Solution
A method using a generative adversarial network (GAN) to convert design patterns into simulated SEM images, adding noise to these images, and training a denoising model with both simulated and captured SEM images to enhance pattern coverage and reduce retraining requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If extensive training data from patterned substrates is used for denoising models, then model accuracy is improved, but metrology time and computing resources increase significantly
Solution Approach 1:
The patent creates simulated SEM images by copying and transforming design patterns through a generative model, rather than capturing real SEM images of patterned substrates. This copying approach provides abundant training data without requiring actual metrology measurements, thus improving model accuracy while avoiding the time cost of extensive substrate scanning.
Solution Approach 2:
The generative model is trained offline in advance to learn the mapping between design patterns and real SEM images. This preliminary action creates a simulated image generator that can produce training data on-demand without requiring actual substrate measurements during model development, significantly reducing metrology time while maintaining data quality.
2Measurement precision
If extensive training data from patterned substrates is used for denoising models, then model accuracy is improved, but computing resources increase significantly
Solution Approach 1:
Instead of collecting and processing large volumes of real SEM images from diverse patterned substrates, the patent copies design patterns through a generative model to create simulated images. This approach provides unlimited training data with minimal computing resources, as the simulated images are generated algorithmically rather than requiring storage and processing of extensive real image datasets.
Solution Approach 2:
The generative model performs preliminary learning offline to capture the statistical characteristics of SEM images. Once trained, it can generate training data rapidly without requiring heavy computing resources during the actual denoising model training phase, as the complex pattern recognition has already been performed in the preliminary training stage.
3Adaptability or versatility
If design patterns are converted to simulated SEM images using GAN, then pattern coverage is improved, but training complexity increases
Solution Approach 1:
The generative adversarial network serves multiple functions: it learns the mapping from design patterns to SEM images, generates diverse training data for various pattern types, and provides a unified framework for creating simulated images across different design patterns. This multi-functionality achieves comprehensive pattern coverage while consolidating the training process into a single versatile model rather than requiring separate processing for each pattern type.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Described herein is a method for training a denoising model. The method includes obtaining a first set of simulated images based on design patterns. The simulated images may be clean and can be added with noise to generate noisy simulated images. The simulated clean and noisy images are used as training data to generate a denoising model.